This paper introduces a novel method to enhance price trend trading strategies using image-induced importance (triple-I) weights derived from deep learning analysis of price chart images. The authors design a ResNet 'trader' that extracts general price patterns from historical price-volume images and generates localization maps via smooth Grad-CAM, which assign importance values to each spatial location for predicting upward or downward price movements. These importance weights are then used to compute a weighted moving average of existing trading signals (e.g., momentum, reversal, trend, and formulaic alphas), amplifying signals at more informative time points. Empirical analysis on the Chinese A-share market (2014-2023) shows that the TWMA method significantly improves portfolio performance: one-day Sharpe ratios increase by at least 219% for momentum and reversal strategies, and by 27-73% for trend and alpha strategies, while also reducing turnover. The method's robustness is confirmed across different network depths, image structures, and stock sizes. Furthermore, the triple-I weights enable effective transfer learning: time-scale transfer allows high-frequency models to enhance long-term (one-month) portfolios, and non-technical transfer improves news-based trading strategies by de-noising sentiment scores. Finally, the TWMA method is validated on 7,846 trading rules, consistently outperforming original and CNN-based signals. This work opens the 'black-box' of deep learning in technical analysis, providing a practical tool for traders to enhance their strategies using interpretable image-derived importance weights.